(14) n-vm

.config/nextest.toml

@@ -136,8 +136,29 @@ store-failure-output = true
 # filter = 'test(flaky test name here)'
 # retries = 3
 
+# Every package that boots a guest, not just the one that owns the harness.
+#
+# A VM claims a 1GiB hugepage for its guest memory and the host pool is small -- two pages on the
+# machine this was found on. Nothing arbitrates that pool, so tests that boot concurrently race for
+# it and the losers fail with "the 1073741824-byte hugepage pool has 0 free page(s)". That is a
+# resource conflict wearing the costume of a flaky test, and it made ten of `dataplane-n-vm`'s
+# eighteen integration tests fail in a whole-workspace run while all eighteen passed with
+# `--test-threads 1`.
+#
+# The list is `#[n_vm::test]`'s users. It has to be maintained by hand, which is the weakness of
+# fixing this here: a package that starts booting guests and is not added to it fails in a way that
+# looks like its own bug. The durable fix is for `n-vm` to acquire the pool itself and wait rather
+# than fail -- a test harness that only works when the consuming repo's runner config is right is
+# not one that can be handed to another repo.
 [[profile.default.overrides]]
-filter = 'package(dataplane)'
+filter = '''
+    package(dataplane)
+  + package(dataplane-n-vm)
+  + package(dataplane-n-it)
+  + package(dataplane-hardware)
+  + package(dataplane-interface-manager)
+  + package(dataplane-mgmt)
+'''
 platform = 'cfg(unix)'
 test-group = 'vm'
 

.github/workflows/dev.yml

@@ -195,6 +195,22 @@ jobs:
       USER: "runner"
       # The `just` action runs under `set -u`, so this stays defined.
       JUST_VARS: ""
+      # These runners are containers on bare metal talking to the host's
+      # Docker daemon, so the daemon cannot see this container's /nix. Measured
+      # on a runner: /nix/store holds 7386 entries here and 0 as the daemon
+      # sees it, while `_work` is the same block device at the same path in
+      # both namespaces. So the closure the container mounts is exported to a
+      # directory under `_work`, where the two agree. `setup-roots` fills it;
+      # `n-vm` reads the same variable and rewrites its bind mount sources,
+      # leaving the targets alone. See `n_vm_host_share` in the justfile.
+      #
+      # Beside the checkout rather than inside it. Two reasons: the workspace
+      # is served to the guest over virtiofs, and -- the one that actually
+      # bites -- `default.nix` filters `./.` as its source, and the markdown
+      # and C-header filters would sweep a few GiB of copied store into every
+      # build. `runner.temp` would be the natural home but the `runner`
+      # context is not available to a job-level `env`.
+      N_VM_HOST_SHARE_DIR: "${{ github.workspace }}/../n-vm-share"
     strategy:
       fail-fast: false
       # Each entry gets its own runner; this limits shared-pool occupancy.
@@ -541,6 +557,14 @@ jobs:
 
       - name: "coverage-archive"
         uses: *just
+        env:
+          # Headroom on the VM kill deadline for an instrumented guest, which
+          # is slower than a plain one and writes its coverage profile on the
+          # way out. Cheap insurance rather than a measured requirement: this
+          # is when a hung guest is shot, not a delay anything waits out, and
+          # overrunning it surfaces as "no parseable test verdict from guest"
+          # -- a passing test reported as a broken harness.
+          N_VM_OVERHEAD_SCALE: "3"
         with:
           recipe: "ci::coverage"
           recipe_args: "${{ matrix.profile }}"

.gitignore

@@ -20,3 +20,8 @@ result*/**
 /bin
 sysroot
 devroot
+testroot
+vmroot
+
+# n-vm host share (see `n_vm_host_share` in the justfile)
+.n-vm-share

.semgrep/rules/no-std-sync-direct.yaml

@@ -17,6 +17,17 @@ rules:
         - concurrency/src/quiescent.rs
         - concurrency/src/slot.rs
         - concurrency/tests/
+        # The in-VM test harness is workspace-independent by design: it
+        # depends on no dataplane crate, so that it can go back to
+        # `githedgehog/testn` without a decoupling project first. Taking the
+        # `concurrency` facade would be exactly that coupling, and would buy
+        # nothing -- none of this code is under loom or shuttle, and it is
+        # host-side test infrastructure rather than dataplane.
+        - n-vm/
+        - n-vm-macros/
+        - n-vm-protocol/
+        - n-it/
+        - n-preinit/
     pattern-either:
       - pattern: use std::sync::Arc;
       - pattern: use std::sync::Weak;